EU Study Maps 480 Biological AI Models, Finds Readiness Gaps

Biological AI advances unevenly across research domains

Artificial intelligence models trained on biological data are progressing at dramatically different speeds depending on the availability and quality of training datasets, according to new research from the European Commission's Joint Research Centre.

The study examined 480 biological AI models and found that protein-centric applications—including structure prediction, function annotation, and molecular design—have reached the most advanced stages of development. These models benefit from decades of curated data in repositories like the Protein Data Bank and UniProt, supported by European research infrastructures including the European Molecular Biology Laboratory.

In contrast, single-cell biology remains significantly less developed despite its clinical relevance for applications such as characterizing tumors to predict immunotherapy response. The primary obstacle is limited and poorly standardized data.

Why it matters

The gap between scientific maturity and real-world deployment readiness creates regulatory blind spots and potential biosecurity risks. High-performing models that excel in research benchmarks may not be validated for clinical or industrial use, yet remain publicly accessible for potential misuse in pathogen design or toxin engineering.

The maturity paradox in biological AI

The research introduces what its authors call a "maturity paradox"—models like AlphaFold and ESM3 demonstrate high domain maturity but remain at low-to-mid technology readiness levels. While these models represent the most advanced developments in their research fields, none has undergone certification for clinical or industrial deployment.

The analysis found that no surveyed model has been subject to an integrated readiness assessment that combines scientific validation with governance frameworks and real-world application benchmarking.

Resource disparities shape development

Three factors determine biological AI model development: training data, computational infrastructure, and collaboration patterns.

Training data distribution remains uneven across domains and regions. The United States and European Union host most datasets, followed by China, the United Kingdom, and Switzerland. Protein models increasingly rely on synthetic data from the AlphaFold Database rather than exclusively curated, experimentally derived data.

Computational resources create a significant gap between industry and academia. Industry developers typically train models on larger datasets with more hardware and longer training times. Europe maintains sufficient high-performance computing capacity through the European High Performance Computing Joint Undertaking and its supercomputing network, recently expanded through AI Factories.

Collaboration patterns reveal that academia participates in developing 85% of surveyed models, while industry participates in nearly 40%. However, only 17% of industry-only developed models release training code, suggesting a trend toward proprietary development.

Geographically, intra-EU collaboration lags behind EU partnerships with the US, China, and the UK. Among the global top 20 model developers, the Technical University of Munich is the only EU representative.

Policy recommendations

The report outlines four priorities for EU action: broadening support for emerging research areas like single-cell biology and multimodal architectures; strengthening biological data infrastructure through better coordination and quality assessment; supporting European biological AI foundation models as public goods while encouraging strategic intra-EU collaboration; and developing frameworks that assess both scientific maturity and technology readiness with clearer regulatory pathways.

The findings were published in the JRC report "Artificial Intelligence for Biology: Capabilities, Readiness, and Policy Implications," first reported by the Joint Research Centre on August 20, 2026.